Ask any video a question.
Get back the exact second.
WAVE VSS indexes a stream or recording, then answers plain-language questions with timestamped moments and grounded, cited summaries — hosted, agent-callable, pay-per-call. No NIMs to self-host. No GPU cluster to babysit. One key, three calls: index, query, summarize.
preview · agent-ready · the pipeline runs on WAVE's own GPU rig today; public routing opens when the gateway lights up wave_vss
The catch: there isn't one
This is the Metropolis VSS approach — chunk the video, caption every chunk with a vision model, embed, rerank, compose a grounded answer — except you self-host none of it. Three calls: index a video, query it, summarize it. Call with a Bearer key and pay per call via x402 — an unpaid call gets a 402 challenge; pay and retry, no subscription. The heavy vision model runs on WAVE's own GPU rig and bursts to cloud GPU when it has to. This edge spoke just carries your key through untouched — it never reads, mints, or strips it.
Where it composes
VSS is agent-callable by design: an agent discovers the spoke at /llms.txt and /.well-known/wave.json and calls it unattended — same Bearer key, same x402 wallet, same meter as every other product on the WAVE Protocol Plane. Video intelligence stops being a project you stand up and becomes one composable call.
1. Index
POST /api/v1/vss/index — point it at a stream, feed, or recording.
2. Query
POST /api/v1/vss/query — ask in plain language, get back timestamped moments.
3. Summarize
POST /api/v1/vss/summarize — a grounded summary, cited to the moments it's built from.